Articulate 0.3.0

What changed

• Conversations now captures meetings from any app using microphone and system audio. Separate Discord participant audio remains an optional source when its companion is connected.
• Conversations, Notetaker, and Dictation have separate workspace views. Call transcripts show clearer speaker turns and overlap.
• An optional Nemotron 3.5 ASR download provides live previews and checks short completed utterances for unexpected writing-system switches. Qwen remains the final recognizer, and calls keep automatic language detection with a per-call preference for ambiguous words.
• Call notes can use speaker attribution, turn times, overlap, and detected sound or tone cues. Cues are uncertain observations and are not treated as proof of intent or agreement.
• App and Vencord companion update notices are visible throughout the workspace. The native Discord adapter has connection and loss handling improvements.

Before you update

The optional Nemotron speech model is downloaded separately. Existing models, transcripts, notes, and preferences remain on your device. If you use the custom Vencord companion, update it from Settings and restart Discord when prompted.

Windows packages are unsigned. See the README for installation and model requirements.

Articulate 0.2.0

Articulate 0.2.0

Articulate now has a new Tauri and React desktop interface for dictation, calls, and spoken notes.

What's new

• A redesigned workspace with smoother transitions, live transcripts, and editable dictations.
• Spoken notes and a quick-note shortcut, with independent dictation sources organized into topic notes.
• Local note generation and descriptive call titles, plus automatic speech cleanup and remembered spelling corrections.
• A transcription-language selector for dictation, calls, and spoken notes. Choose English in Settings > Audio to reduce unwanted language switching, or keep Auto-detect for multilingual conversations.
• Optional local sound cues for laughter and other audible events. Tone estimation has been removed.

Fixes and improvements

• More reliable automatic Discord capture, companion pairing, installation status, and separate participant audio.
• Better handling of delayed audio, reconnects, packet loss, and back-and-forth speaker turns.
• Discord usernames for participant identification, plus visual and audible capture feedback.
• Improved silence filtering and clearer interruption notices.
• Transcript-specific editing and notes state, vocabulary context merging, and a cleaner recording overlay.

Windows downloads

Use the setup executable for installation, or extract the portable ZIP and run Articulate.exe. Models download separately from Settings. Windows requires the Microsoft Edge WebView2 runtime; the installer links to the Microsoft download page when it is missing.

Existing local recordings and settings are retained. If you use the Vencord companion, update it in Settings and fully restart Discord to load the new companion. Saved tone estimates are omitted when recordings are opened or exported.

Sound-event detection can miss reactions or label them incorrectly. Choosing a transcription language reduces unwanted language switching but does not eliminate recognition errors.

Articulate 0.1.3

Articulate 0.1.3

A Notetaker workspace, local summaries and cleanup, saved-content search, and private usage Insights.

• Notetaker. Create personal notes without recording, write during a call, and return to saved conversations. Generous reading and writing areas keep your notes, transcript, highlights and summary separate.
• Local summaries. An optional Qwen3.5 4B model generates facts, decisions and action items from completed calls and dictations. Review exact source excerpts before saving. Longer transcripts are processed in sections, so later changes to a plan still need review.
• Find saved information. Search words or phrases throughout retained call transcripts, dictations and personal notes. Open the original source from a result.
• Insights. See dictated words, measured pace, vocabulary and cleanup changes, usage by app, and daily activity. Statistics use retained local history; older recordings may lack timing and correction measurements.
• Cleanup previews. Polish short dictations with Clear, Professional or Casual styles. Compare the original and preview before applying. Text already inserted in another app stays untouched.
• Reliable downloads. Cleanup and summary model downloads resume after interruptions, retry stalled transfers and verify checksums. The previous total-body timeout is fixed.
• Discord auto-start. Connection status distinguishes participant names from separate audio readiness. Automatic capture can load an existing selected speech model, waits for participant audio readiness and retries startup failures. Manual captures remain under your control.

Models download separately. The cleanup editor is about 852 MB including its runtime; the summary model is about 3.16 GB on first setup. Neither is bundled in the installer. The fast Rust cleanup remains the live dictation path. Summary generation and cleanup previews require idle recording.

Generated text needs review. Source citations confirm provenance, not interpretation. Highlights remain quoted transcript excerpts. Saved-content search is text search, not semantic question answering. Calendar integration and cross-device sync are not included. Native Discord participant audio still needs real desktop-call validation. Windows builds are unsigned.

Existing models, vocabulary, preferences, transcripts and notes are preserved when updating. Download the installer or extract the complete portable ZIP.

Articulate 0.1.2

Articulate 0.1.2

A cleaner workspace, easier dictation, and more control over vocabulary and notes.

• Hold to talk. Hold Ctrl+Alt+Space and release to finish, or double-press for hands-free dictation and press once to finish. Keys and activation mode are customizable. Recording start/stop sounds can be turned off. Hold mode inserts into your app after finishing; toggle mode supports live insertion in compatible fields.
• A clearer interface. Persistent sidebar navigation, compact capture controls, readable call transcripts, cleaner History rows, and grouped Settings. Editor Save actions remain visible in smaller windows.
• Vocabulary that builds on corrections. Matching spellings combine their context rules instead of creating duplicate entries. Review and edit the contexts and capitalization rules. Learning notices distinguish updates from new entries, and Undo restores the previous rule.
• More useful review. Choose individual suggested note passages before saving. Notes preserve exact transcript sentences and their speaker-turn time ranges. Vocabulary review shows the original and saved spelling for explicit selection; it does not automatically rewrite live text.
• Explicit Discord audio selection. Selecting the Vencord companion requires its separate participant-audio connection. Capture waits for that connection instead of silently falling back to mixed output. Mixed capture remains available for other call sources.

The bundled Assort model weights are unchanged. Their suggestions still need review. Large speech and speaker models download separately; the small Assort models are included.

Native Discord participant audio remains experimental: real desktop-call validation is still pending. Mixed output cannot reliably separate simultaneous speakers. Windows builds are unsigned and may show an unknown-publisher warning.

Download the installer or extract the entire portable ZIP. Existing models, vocabulary, settings, and saved sessions are preserved when updating.

Articulate 0.1.1

Articulate 0.1.1 is a Windows testing release with included local classifiers, Discord connection fixes, and more reliable live dictation.

Download

• Articulate-0.1.1-windows-x86_64-setup.exe: recommended per-user installer.
• Articulate-0.1.1-windows-x86_64.zip: portable version. Extract the whole folder before opening Articulate.exe.
• SHA256SUMS.txt: checksums for both downloads.

Existing users can check for this release in Settings > App updates. Finish recording before installing. Saved sessions, settings and downloaded models are kept.

Changes

• Pretrained Assort models for meeting highlights and contextual vocabulary review are included. No model paths or separate classifier installation are needed. Suggestions remain reviewable and do not automatically rewrite your text.
• Fixed a Windows connection-handling issue that could repeatedly disconnect the Vencord companion.
• Discord connects automatically after configuration. An optional setting starts a transcript on a confirmed call join and finishes it on leave.
• The custom Vencord build includes the native separate-participant audio adapter by default. Supported Discord voice-module versions are checked before use.
• Improved live insertion and revision in Discord's plain-text message field, including empty-editor recovery and accented characters.
• Refined transcript, notes, history and settings layouts.

First run

Download the speech model from Settings > Audio & models. Qwen3-ASR 1.7B Q8 is about 2.19 GB. The optional acoustic speaker model is about 237 MB. These large models are not bundled; the small Assort classifiers are.

Windows 10/11 x64 and 16 GB RAM are recommended. CPU-only operation is supported, with optional Vulkan acceleration. Press Ctrl+Alt+Space to start and finish dictation; the shortcut is customizable.

Testing limits

• Separate Discord audio has passed synthetic ABI and transport tests; real desktop-call validation is still pending. Mixed output cannot reliably separate overlapping voices. The app identifies the capture source in use.
• Discord emoji and other embedded editor objects can still pause live insertion. The transcript remains available in Articulate.
• Assort uses synthetic English training examples and can misclassify passages. Review suggestions. Notes contain quoted highlights, not generated narrative summaries.
• Recognition accuracy and speed depend on speech and hardware. Broader long-session and hardware testing is ongoing.
• The Windows binaries are unsigned. macOS, mobile keyboards and automatic device sync are not included.

Audio and text are processed locally. Large model downloads, explicit update checks and Discord avatar retrieval require network access. Audio is not archived. Please use synthetic or non-sensitive examples in public feedback.

Licensed under AGPL-3.0-or-later. Corresponding source and build scripts are available in the v0.1.1 source archives and tag. Third-party notices are included in both binary distributions.

Articulate 0.1.0

Articulate's first Windows testing release: local dictation, call transcripts, vocabulary corrections, vocal text templates, and saved history.

Download

• Articulate-0.1.0-windows-x86_64-setup.exe: recommended installer, no administrator access required.
• Articulate-0.1.0-windows-x86_64.zip: portable version. Extract the whole folder before opening Articulate.exe.
• SHA256SUMS.txt: checksums for both downloads.

First run

Open Settings and download the recommended speech model (about 2.19 GB). Calls also need the separate speaker model (about 237 MB). No model weights are bundled. Windows 10/11 x64 and 16 GB RAM are recommended; CPU-only operation is supported, with optional Vulkan acceleration.

Press Ctrl+Alt+Space to start and finish dictation. The shortcut can be changed in Settings. Updates can be checked and downloaded from Settings; recording and pending history saves must finish before installation.

Included

• Live dictation into supported text fields, editable vocabulary, and text/template shortcuts.
• Microphone and call-output capture, revisable call text, speaker labels, quoted notes, and export.
• Local History with search, rename, saved notes, and deletion.
• Optional Discord speaker metadata through the standard client debugger or custom Vencord companion source.

Testing notes

The Windows build is unsigned. Obtain it from this repository. Recognition can make mistakes, especially with noise, names, overlap, and short utterances. Discord currently supplies speaker activity, not isolated participant audio. The Vencord companion is a custom source plugin and requires a Vencord source build. macOS, mobile keyboards, cloud sync, and generative meeting summaries are not included.

Audio is not saved. Transcripts, notes, preferences, and downloaded models remain on the device and survive uninstall. Please use synthetic or non-sensitive examples in public bug reports.

Licensed under AGPL-3.0-or-later. Corresponding source and build scripts are in this release's source archives and the v0.1.0 tag.